Prompt in, answer out. Once.
The model the internet was built on: a prompt goes in, a model answers, the exchange ends. It cannot open a tool, check its own work or come back with a question — every next step is yours to type.
How an agent works
A loop, not a request.
The internet was built on request → response: a prompt goes in, an answer comes out, and the model never acts. An agent runs a loop — plan, act, review, ask, report, grow — until the job is done, on a harness that executes and records while the agent decides.
Plan, act, review — around until the job is done.
An agent plans, acts — calling tools in batches — reviews what came back, asks you when it is unsure, reports, and grows from what happened. Around again until the job is done, every step streamed live.
The harness
The run loop around the model. The agent decides; the harness executes and records. It
- hands the model its tools and their guides
- executes each tool call with your credentials
- keeps the journal and the insight after each batch
- enforces the turn contract — Lite, Standard, Tools
- carries the clock: the temporal header
- compacts on overflow and meters every call
- stops the moment you press Stop
-
Plans, acts, reflects
A turn opens with a plan and a preamble of what it is about to do. Tools run in batches; after each one the agent reflects, may write a finding, and may replan.
-
Asks when unsure
ask_user pauses the run with choices or a short form. You answer; the run picks up where it stopped. No guessing on your behalf.
-
Learns over time
A per-tool Do / Don’t ledger of what worked and what failed on your account; a Growth Reviewer proposing new Skill and Role versions with evidence and a diff; training against a plan you review first.
-
Knows the time
A temporal header gives every iteration the current time in your zone, so “tomorrow” and “last quarter” mean what you mean.
-
Runs your workflows as tools
A saved workflow becomes one tool the agent can call — a whole pipeline as a single step, executed as an isolated Job.
-
Reads your knowledge
A dedicated search subagent reads whole documents mid-run, follows references and cites what it used.
Design a custom agent → Agent, defined Agent mode Back to the homepage